The biggest mistake in AI consulting isn’t failing to use AI—it’s using AI where it isn’t needed. The solution pyramid is the decision-making framework the consultant uses in every diagnosis: each level has its own cost, risk, and suitability profile. Move up only when the problem requires it.
Illustrative diagram — the solutions pyramid: move up only when the problem requires it.
🧱 Base: deterministic workflows — cheap, fast, reliable
The base of the pyramid is where most solutions should sit. Deterministic means predictable: given the same input, always the same output. Rules, fixed automation, API integrations, specialized SaaS, structured database.
✅ What belongs in the foundation
- •Fixed business rules — “if order is over R$ 500, shipping is free” → if/else, not an LLM.
- •Repetitive task automation — forms, Zapier/n8n integrations, webhooks.
- •Specialized SaaS — there’s a ready-made product that solves exactly that problem.
- •Database and reports — SQL queries, dashboards, filters.
💡 Why the foundation is underestimated
A deterministic solution doesn’t impress in a presentation. But it’s the cheapest to build, the easiest to maintain, and the hardest to break. A client who saves 20h/week with a simple automation is a happy client — whether it uses AI or not.
🤖 Middle: AI workflows — more power, more cost, more points of failure
The middle floor is where AI comes in surgically: one process step uses an LLM; the rest of the workflow remains deterministic. When applied well, this is the highest-ROI layer because it combines AI’s power with the stability of fixed automation.
✓ Ideal cases for the middle
- ✓Classify incoming emails before routing them
- ✓Extract data from unstructured documents (PDFs, contracts)
- ✓Generate a standard response draft for human triage
- ✓Summarize long transcripts into action items
📊 What Changes vs. the Baseline
- →Cost: LLM APIs charge per token — costs scale with volume
- →Latency: LLM calls add 1–5s to the workflow
- →Failures: hallucination, timeout, output in the wrong format
- →Oversight: someone needs to monitor output quality
🛰️ Top: AI agents — maximum capability and risk
Agents are systems with autonomy to decide and act: they receive a goal, choose tools, execute steps, observe results, and iterate. The power is greatest — and so is the risk. Agent projects have the longest delivery times and the highest maintenance costs.
🛰️ When agents make sense
- •Complex multi-step task where the steps depend on the previous result and aren’t known at design time.
- •Use of tools — the agent needs to search the web, query databases, call APIs, or run code.
- •Refinement loop — the agent evaluates its own output and iterates until it meets the quality criteria.
⚠️ The risks at the top
- Cascading error: a wrong decision in step 2 contaminates all subsequent steps.
- Unpredictable cost: long loops multiply tokens and tool calls.
- Difficult debugging: emergent behavior is difficult to track and reproduce.
- User trust: when the agent makes an unexpected mistake, the damage to trust is greater.
📐 The axes: cost, time, and risk increase as you move up
The pyramid isn’t just about technical complexity. It’s about three business dimensions the client needs to understand before approving any budget: cost, delivery time, and operational risk.
📊 The three axes on each floor
| Axis | 🧱 Base | 🤖 Middle | 🛰️ Top |
|---|---|---|---|
| Build cost | Low | Medium | High |
| Operating cost | Predictable | Scale with volume | Unpredictable |
| Delivery time | Days to weeks | Weeks | Months |
| Failure risk | Low | Medium | High |
| Maintenance | Simple | Moderate | Complex |
💡 Use the table with the client
Showing these axes before making any proposal is one way to align expectations. A client who chooses the top option knowing the cost and risk is a prepared client — and a partner when things get difficult.
🪜 The golden rule: start at the base, move up only when necessary
The rule is simple and rarely followed: use the lowest level of the pyramid that solves the problem. If the baseline solution works, great. If you need AI for one step, move up one level. If you need an agent, move up two. But never move up without trying the level below first.
✓ Apply the rule
- ✓Assesses whether simple automation can solve it before proposing an LLM
- ✓Proposes an agent only when autonomous, multi-step work is necessary
- ✓Documents why you chose level X instead of X-1
- ✓Starts with the smallest scope that proves value
✗ Breaking the rule
- ✗Proposes an agent for a task a simple rule can handle
- ✗Uses an LLM for binary classification with fixed criteria
- ✗Chooses the level based on the presentation’s “wow factor”
- ✗Doesn’t test the simplest solution because they judge it inadequate
🧪 Examples by level — the same problem at each layer
The same business problem can be solved on all three tiers — with very different cost, speed, and robustness profiles. Seeing all three sides helps the client choose consciously, not hype.
Problem: "I want to automate answering frequently asked questions on WhatsApp"
🧱 Base (deterministic)
WhatsApp Business API options menu + spreadsheet FAQ + keyword-based conditional flow. Cost: R$ 500-2k/month. Works for 80% of questions.
🤖 Middle (AI workflow)
Core workflow + LLM that understands variations in the question and generates an answer based on the knowledge base (RAG). Cost: R$ 2-8k/month. Covers 95% of questions.
🛰️ Top (agent)
An agent that checks the CRM, verifies an order, negotiates returns, and escalates to a human with full context. Cost: R$ 15-50k build + R$ 5-15k/month. Needed only for complex customer service with access to customer data.
💡 The value of comparison
Showing the three options with costs and trade-offs is what sets the consultant apart from the salesperson. The salesperson only shows the option they sell. The consultant shows all three and helps the client choose the right one for their context — today.
🎒 Module summary
Next module:
1.4 — When NOT to use AI—and how that builds trust